Assessing the prognostic power of prediction rules: A Users’ Guide to the medical literature

Article type
Authors
Alba AC1, Agoritsas T2, Walsh M2, Hanna S2, Iorio A2, Devereaux P2, McGinn T3, Guyatt G2
1Toronto General Hospital - University Health Network - Toronto - Ontario
2Department of Clinical Epidemiology and Biostatistics, McMaster University, Hamilton, Ontario
3The Feinstein Institute for Medical Research, Hofstra Northwell School of Medicine, Hempstead, New York
Abstract
Accurate prognostic information is fundamental to optimal clinical care. Patients and clinicians can use their intuition and average risk from observational studies to estimate prognosis. The best approach to assess patient prognosis, however, relies on models that simultaneously consider a number of prognostic factors and provide an estimate of patients’ absolute risk of an event. Such predictive models or rules should be characterised by adequate discrimination - differentiating patients who will have an event from those who will not - and adequate calibration - ensuring accurate prediction of absolute risk. We have developed a Users’ Guide to understanding the available metrics for assessing discrimination, including the area under the receiving operator curve and c-statistic; calibration, through comparison of observed and predicted risks; and the relative performance of two different models by risk-reclassification analysis. This presentation will use real-world examples to demonstrate concepts and apply results from studies evaluating predictive models. This guide complements an existing Users’ Guide that addresses the development and validation of predictive models. Together, these guides will allow clinicians to make optimal use of existing prediction models.